Gut Microbiota and Body Weight in School‐Aged Children: The KOALA Birth Cohort Study

Abstract

Objective

This study aimed to examine the intestinal microbiota composition of school‐aged children in association with (over)weight.

Methods

The fecal microbiota composition of 295 children was analyzed using the Human Intestinal Tract Chip. Anthropometric outcomes (overweight [BMI  ≥  85th percentile], age‐ and sex‐standardized BMI and weight z scores) were measured at 6 to 7 years of age, and elastic net was used to select genus‐like bacterial groups related to all anthropometric outcomes. Subsequently, multiple linear and logistic regression models were used to model associations between selected bacterial groups and anthropometric measures while controlling for confounders.

Results

Prevotella melaninogenica, Prevotella oralis, Dialister, and uncultured Clostridiales II (UCII) accounted for 26.1% of the variation in microbiota composition. Several bacterial groups were inversely associated with the anthropometric outcomes: Sutterella wadsworthensis, Marvinbryantia formatexigens, Prevotella melanogenica, P oralis, Burkholderia, uncultured Clostridiales II, and Akkermansia, while Streptococcus bovis was positively associated with overweight. Microbial diversity and richness, and Bacteroidetes to Firmicutes ratio, were not significantly associated with any of the outcomes.

Conclusions

In the largest population‐based study on childhood gut microbiota and body weight so far, both new and previously identified bacterial groups were found to be associated with overweight. Further research should elucidate their role in energy metabolism.

Introduction

Whereas excessive energy intake and insufficient physical activity are the main drivers of childhood overweight and obesity, recent research has suggested that other factors such as the gut microbiota may also be involved 1. The gut microbiota is highly diverse in composition and plays an important role in human physiology, metabolism, nutrition, and immune function 2. Evidence from some human, but mostly animal, studies has suggested that the gut microbiota may contribute to the development of overweight via mechanisms involving increased energy harvest 3, regulation of host metabolism 4, and the activation of innate immunity 5. In humans, obesity has been associated with lower gut microbial richness and diversity 6, 7 and, in some studies, a reduced Bacteroidetes to Firmicutes (B:F) ratio 8, 9, whereas others have reported the opposite 10, 11 or no association at all 12, 13. Associations between specific bacteria, e.g., Bacteroides fragilis 14, Bifidobacterium species, Staphylococcus species 15, Akkermansia muciniphila 6, Faecalibacterium prausnitzii 16, or between specific archaea, e.g., Methanobrevibacter smithii 17, and obesity in humans have also been reported, although the identified microbial groups vary greatly between studies.

These inconsistent findings might be attributable to the use of miscellaneous methods to assess the gut microbiota, often enumerating specific taxa rather than using broad 16S ribosomal RNA (rRNA) gene surveys or metagenomics. Moreover, studies vary greatly in the populations considered, their designs, and the degree of control for potential confounding factors such as lifestyle and diet 1. Most studies have compared extreme categories of lean adults and adults with obesity, while the gut microbiome is less well studied in healthy children of primary school age. We therefore aimed to investigate the, probably more subtle, relation between gut microbiota composition and body weight in a large group of 295 well‐phenotyped, healthy, school‐aged children.

Methods

Subjects and study design

The current study is a cross‐sectional analysis conducted within the observational KOALA Birth Cohort Study. The design of the KOALA study has been described in detail elsewhere 18. Briefly, a total of 2,834 pregnant women were recruited at 34 weeks of gestation, from October 2000 until December 2002. Healthy pregnant women with a conventional lifestyle (N = 2,343) were recruited from an ongoing cohort study on the etiology of pregnancy‐related pelvic girdle pain in the Netherlands 19. An additional 491 pregnant women with self‐reported alternative lifestyles were recruited through organic food shops, anthroposophist doctors and midwives, Steiner schools, and dedicated magazines. This latter group of women was considered to have an alternative lifestyle that could involve dietary habits (vegetarian, organic food choice), child‐rearing practices, vaccination schemes, and/or use of antibiotics 18.

A subgroup of 1,204 parents was asked for consent for a home visit for anthropometric measurements and to collect a single fecal sample from the child at the age of 6 to 7 years. This subgroup was composed of participants who had home visits for blood collection from the mother during pregnancy and/or the child at age 2 and who were still active participants (Supporting Information Figure S1). Fecal samples were obtained for 669 children. Exclusion criteria for the current study were as follows: prematurity (< 37 weeks of gestation), twins, abnormalities linked to growth (such as Down syndrome, Turner syndrome, tetralogy of Fallot, multiple disabilities, and cystic fibrosis), fecal samples with transport times exceeding 3 days, or lack of data on dietary intake. A total of 295 children were ultimately included in the present study, all white. Written informed consent was given by all parents, and the study was approved by the Medical Ethics Committee of Maastricht University and the National Ethical Committee for Medical Research. Data collection was performed in accordance with relevant guidelines and regulations.

Data collection and outcome measures

Fecal collection

Fecal samples of the children were collected by the parents at home upon receipt of a feces tube with a spoon attached to the lid (Sarstedt, Nürmbrecht, Germany) together with instructions for collection. After collection, samples were sent to the laboratory by mail. Upon arrival, samples were diluted 10‐fold in Peptone/water (Oxoid, Altrincham, UK) containing 20% (vol/vol) glycerol (Merck, Darmstadt, Germany) and stored at −80°C until further analysis.

Fecal DNA isolation and microbiota profiling

DNA isolation from fecal samples has been described in detail elsewhere 20. Briefly, DNA was isolated using a combination of repeated bead beating and column purification. DNA concentration and purity were assessed with a NanoDrop 1000 spectrophotometer (Thermo Fisher Scientific, Wilmington, Massachusetts). DNA was then stored at −20°C pending microbial analysis, which was performed using a previously described and benchmarked 21, 22, custom‐made phylogenetic microarray, the Human Intestinal Tract Chip (Agilent Technologies, Santa Clara, California), which contains a duplicated set of 3,631 probes targeting the 16S rRNA gene sequences of more than 1,000 intestinal bacterial phylotypes 21. After DNA extraction, the full‐length 16S rRNA gene was amplified, followed by in vitro transcription and labeling of the resultant RNA with Cy3 and Cy5 (GE Healthcare, Amersham, UK) before hybridization to the array. Each sample was hybridized at least twice to ensure reproducibility, and only duplicate hybridizations with a Pearson correlation > 0.98 were considered for further analysis. Microbiota profiles were generated by preprocessing the probe‐level measurements with minimum–maximum normalization and Robust Probabilistic Averaging probe summarization 23, 24 into the following three phylogenetic levels: level 1, defined as order‐like 16S rRNA gene sequence groups; level 2, defined as genus‐like 16S rRNA gene sequence groups (sequence similarity > 90%); and level 3, defined as phylotype‐like 16S rRNA gene sequence groups (sequence similarity > 98%). Here, we primarily focus on the genus‐level variation referred to as species and relatives (“et rel.”). The log10‐transformed signals were used as a proxy for bacterial abundance. B:F ratio was calculated by dividing the total hybridization signal intensity for Bacteroidetes by that for Firmicutes. Diversity of the microbiota was quantified using the Shannon diversity index based on nonlogarithmic oligo‐level signals as implemented by the R package vegan (Oksanen J, Blanchet F, Kindt R, et al. vegan: Community Ecology Package. R package version 2.0‐10. 2013.) Probes were counted in each sample to measure richness by using an 80% quantile threshold for detection.

Anthropometric outcomes

To obtain highly reliable data, height and weight were measured during home visits at the age of 6 to 7 years by trained research assistants, with the children wearing only their underwear. Height (in millimeters) was measured using a portable stadiometer (Leicester height measure, Marsden Weighing Group, Rotherham, UK), and weight (in grams rounded to 100 g) was measured using a digital scale (HE‐5; CAS Corp., East Rutherford, New Jersey). BMI was calculated as weight divided by height squared. Because BMI values as such cannot be meaningfully interpreted in growing children, BMI and weight and height measurements were converted into age‐ and sex‐specific z scores using the national Dutch Growth Study as a reference population 25. BMI z scores were used as both continuous and dichotomous outcomes: without overweight versus with overweight (BMI z score ≥ 1.04, corresponding to the 85th percentile) 26.

Confounders: Other phenotypic and lifestyle data

All collected data that could influence the relation between the childhood gut microbiota and body weight were considered as confounders (Supporting Information Table S1). These data were collected at several time points throughout the study. At 14 and 34 weeks of gestation, pregnant women received questionnaires regarding, among other questions, family size, prepregnancy height and weight, weight gain during pregnancy, and maternal smoking. Two weeks after childbirth, data were collected from obstetric reports, and questionnaires were completed by the mothers to obtain information on gestational age, birth weight, and sex of the child. Food frequency questionnaires were filled out by the parents to report the dietary habits of their children at the age of 5.0 ± 0.6 years (mean ± SD). Information on physical activity was obtained from parents, and antibiotic use prior to the home visit was recorded.

Statistical analysis

Using statistical software package Canoco 5 27, redundancy analysis, a multivariate canonical ordination analysis method, was performed to determine how much variation in microbiota composition was explained by the anthropometric outcomes and selected confounders. As described below, this was followed by elastic net‐based selection of relevant bacterial groups, and multivariable linear and logistic regression, to obtain direct, independent associations between selected bacterial groups and anthropometric outcomes after adjustment for a set of a priori‐selected confounders (Supporting Information Table S1).

Selection of potentially relevant bacterial groups

Elastic net is a method that utilizes the ridge ( = 0) and lasso ( = 1) penalties to perform both shrinkage and automatic variable selection simultaneously. It also addresses the problem of multicollinearity by encouraging a grouping effect, whereby strongly correlated bacterial groups are kept in the model. The data were randomly divided into a training set with 197 observations and a test set with 98 observations (Supporting Information Figure S2). Model fitting and tuning parameter (λ and s) selection was done on the training set by performing a 10‐fold cross validation. The first tuning parameter (λ) plays a role in variable selection, whereas the second (s) captures correlated predictors at the same time. The chosen grid values were 0, 0.01, 0.1, 1, 10, and 100 for λ and 0 to 1 dividing on a scale of 0.1 for s, as previously suggested 28. The most parsimonious model was obtained using the combination of the two tuning parameters corresponding to the smallest mean‐squared prediction error based on the cross validations performed on the training set. Validation of the model performance was then performed by comparing the computed prediction mean‐squared error (for continuous outcomes) on both the training and test set. In the case of the binary outcome (with overweight: yes versus no), validation of the performance of model prediction was done using area under the curve.

In addition, we planned a priori to examine the association between anthropometric outcomes and bacterial groups that exhibited a strong bimodal abundance distribution, calculated using potential analysis 29 with the early warnings R package 30. These bimodal bacterial groups were previously described by Lahti et al. as potential tipping elements in the ecosystem 31 and might serve as indicators of the community state and its link with (over)weight. We selected bacterial groups that showed consistent evidence for multimodality with bootstrap support ≥ 68%.

Regression analysis

First, we performed unadjusted regression models (Model 1) to analyze the association of each individual selected genus‐like bacterial group, bimodal bacterial group, microbial richness and diversity, and B:F ratio with the outcomes (linear regression for BMI and weight z scores; logistic regression for overweight yes/no). Second, we performed the same analyses while adjusting for confounders considered a priori (Model 2; see Supporting Information Table S1 for the list of confounders). We investigated whether the association between the intestinal microbiota composition and outcomes differed between the two different recruitment channels (alternative and conventional) by including a recruitment group–exposure interaction term in the models. This interaction was significant only for B:F ratio with regard to BMI and weight z scores; the stratified analyses did not, however, provide any statistically significant associations. Finally, we performed multiple regression models including all selected bacterial groups, the bimodal bacterial groups, and archaea (M smithii [yes/no] and Methanosphaera stadtmanae [yes/no]) while adjusting for confounders considered a priori (further referred to as Model 3). Archaea were added because, in our previous study 17, the presence of archaea, specifically M smithii, was associated with childhood weight development.

Maternal weight gain during pregnancy was the only confounding variable with ≥ 5% missing values. Multiple imputations using the Markov chain Monte Carlo method for multiple imputations were performed to assess whether imputed data impacted the outcomes. Results obtained from combined imputed data sets (n = 10) were comparable to those of the original nonimputed data; therefore, final analyses were done without imputation. Analyses were performed in SAS version 9.3 (SAS Institute, Cary, North Carolina) and in R version 3.1.3 (R Foundation for Statistical Computing, Vienna, Austria) using the following packages: elastic net (for continuous outcomes) and glmnet (for binary outcomes). Significance level was set at < 0.05.

Results

General characteristics of the total KOALA population and study population are presented in Table 1. In total, 295 subjects (148 [50.2%] boys and 147 [49.8%] girls) with a mean age of 7.4 years (SD 0.8) were included in the present study. Antibiotic use was limited to 2.4% of the children in the 4 weeks prior to stool collection and 12.4% in the previous year but longer than 4 weeks prior. Postal delivery of the fecal samples took 2.2 days (SD 0.86). A distribution of BMI z scores is presented in Supporting Information Figure S5. The numbers of children in the underweight, overweight, and obesity categories based on BMI z scores ≤ −1.04, ≥  1.04, and ≥  1.65, corresponding to the 15th, 85th, and 95th percentiles in the Dutch reference population, were 45 (15.2%), 27 (9.8%), and 11 (3.7%), respectively.

Table 1.

General characteristics of KOALA Birth Cohort and study population

	KOALA Birth Cohort Study (N = 2,834)a	Study population (N = 295)b
Age at home visit (y)	‐	7.4 ± 0.8
Anthropometric data measured at home visit
Weight z scores	‐	−0.20 ± 0.96
BMI z scores	‐	−0.16 ± 0.89
Overweight (BMI z score ≥ 1.04), n (%)
Yes	‐	27 (9.8)
No	‐	268 (90.2)
Time of last antibiotic course prior to home visit, n (%)
No antibiotic use in the previous year	‐	248 (85.2)
Greater than 4 weeks ago	‐	36 (12.4)
Less than 4 weeks ago	‐	7 (2.4)
Age at fecal sample collection (y)	‐	7.3 ± 0.8
Child’s total physical activity (h/wk)	‐	9.4 ± 4.5
Recruitment group, n (%)
Conventional	2,343 (82.7)	218 (73.9)
Alternative	491 (11.3)	77 (26.1)
Maternal educational level, n (%)c
Low	289 (10.7)	21 (7.2)
Middle	1,060 (39.4)	113 (38.8)
High	1,341 (49.9)	157 (54.0)
Maternal prepregnancy weight (kg)	67.7 ± 13.1	68.3 ± 11.5
Maternal weight gain in pregnancy (kg)	14.3 ± 5.1	14.5 ± 4.8
Place and mode of delivery, n (%)
Vaginal delivery at home	1,187 (44.8)	143 (49.0)
Vaginal delivery at hospital	1,149 (43.4)	120 (41.1)
Cesarean section at hospital	311 (11.8)	29 (9.9)
Maternal smoking in late pregnancy, n (%)
Yes	200 (7.1)	12 (4.1)
No	2,634 (92.9)	283 (95.9)
Total household size	4.3 ± 0.8	4.3 ± 0.8
Gestational age	39.8 ± 5.0	39.8 ± 3.7
Birth weight (g)	3503 ± 512	3,605 ± 466
Breastfeeding duration (mo)	4.7 ± 3.0	6.0 ± 4.4
Sex, n (%)
Male	1,451 (51.2)	148 (50.2)
Female	1,376 (48.6)	147 (49.8)
Child’s dietary intake
Total energy intake (kJ)	6,173 ± 1,285	6,180 ± 1,217
% Energy intake from fats	29.6 ± 4.2	29.7 ± 4.2
% Energy intake from carbohydrates	55.8 ± 5.0	55.7 ± 4.8
Total fiber intake (g)	15.3 ± 4.0	15.6 ± 3.9

^aValues are mean + SD, unless otherwise indicated.

^bTotal may not sum up to 2,834 and 295, respectively, because of missing data.

^cLow: primary school, preparatory vocational, or lower general secondary school; middle: vocational, higher general secondary, and preuniversity; high: higher vocational or academic.

The gut microbiota of the children was dominated by Bifidobacterium and Prevotella melaninogenica, Bacteroides vulgatus, Ruminococcus obeum, and F prausnitzii et rel. (Figure 1). Furthermore, we observed a bimodal distribution of P melaninogenica, Prevotella oralis et rel., Dialister, and uncultured Clostridiales II (UCII), but not for B fragilis et rel. and uncultured Clostridiales I (UCI). A distribution plot for each of the bacterial groups exhibiting bimodality (classified as low or high), as well as their bootstrap support and cutoff or tipping points, is illustrated in Supporting Information Figure S3. The cumulative abundance of these four bimodal groups accounted for 26.1% of the total microbiota variation (Figure 2), while BMI z scores, weight z scores, and overweight explained only 0.5% (P = 0.07), 0.6% (P = 0.04), and 0.5% (P = 0.5), respectively (data not shown).

Figure 1.

Box plots representing the relative abundance of the 21 most abundant taxa (mean abundance > 1%).

Figure 2.

Abundance of bimodal taxa shape overall composition. RDA visualizing microbiota composition of all fecal samples (n = 295) colored by high (black dots) and low (grey dots) abundance of Dialister (left panel) and Prevotella melaninogenica (right panel). Individuals with overweight are represented as squares. The direction of the arrows depicts the abundance of the bimodal bacterial groups as well as their co‐correlating groups. Length of the arrows is a measure of fit for the species. Redundancy analysis (RDA) displays and explains the variation explained in the microbiota, constrained by the predictor variables (bimodal taxa).

Considering the bimodal bacterial groups, a high abundance of UCII was inversely associated with BMI and weight z score (adjusted β −0.22; 95% CI: −0.42 to −0.02 for both; Table 2) and overweight (adjusted odds ratio [OR] 0.28; 95% CI: 0.10 to 0.79; Table 3) in the adjusted analysis. The risk of overweight was lower when abundances of P melaninogenica et rel. (adjusted OR 0.21; 95% CI: 0.07 to 0.68) and P oralis et rel. (adjusted OR 0.20; 95% CI: 0.06 to 0.64) were higher (Table 3). Regarding microbial richness, diversity, and B:F ratio, we found only a borderline significant inverse association between microbial diversity and overweight in the unadjusted linear regression model.

Table 2.

Multiple linear regression results for associations of selected bacterial groups from elastic net, bimodal groups, diversity, richness, and B:F ratio with BMI and weight z scores

					BMI z scores	Weight z scores
Model 1 a , unadjusted β (95% CI) b	Model 2 c , adjusted β (95% CI) c	P value d	Model 1, unadjusted β (95% CI)	Model 2 c , adjusted β (95% CI) c	P value d
Log10 abundance of bacterial groups
Akkermansia			−0.42 (−0.69 to −0.16)	−0.29 (−0.56 to −0.03)	0.030	−0.47 (−0.76 to −0.19)	−0.35 (−0.62 to −0.09)	0.009
Sutterella wadsworthensis et rel.			−0.66 (−1.04 to −0.28)	−0.51 (−0.88 to −0.13)	0.008	−0.47 (−0.89 to −0.05)	−0.30 (−0.68 to 0.08)	0.125
Bacteroides fragilis et rel.			−0.42 (−0.70 to −0.13)	−0.12 (−0.41 to 0.18)	0.437	−0.51 (−0.81 to −0.21)	−0.15 (−0.45 to 0.15)	0.317
Butyrivibrio crossotus et rel.			−0.49 (−0.84 to −0.15)	−0.29 (−0.64 to 0.06)	0.099
Clostridium cellulosi et rel.			0.03 (−0.19 to 0.24)	0.01 (−0.19 to 0.22)	0.897
Bifidobacterium						0.15 (−0.13 to 0.44)	0.06 (−0.20 to 0.32)	0.665
Marvinbryantia formatexigens et rel.						−0.65 (−1.15 to −0.16)	−0.45 (−0.89 to 0.00)	0.050
Streptococcus bovis et rel.						0.14 (−0.10 to 0.39)	0.15 (−0.07 to 0.37)	0.187
Bacteroides vulgatus et rel.						−0.40 (−0.66 to −0.14)	−0.19 (−0.43 to 0.05)	0.120
Bacteroides intestinalis et rel.						−0.10 (−0.45 to 0.25)	0.12 (−0.19 to 0.44)	0.459
Burkholderia						−1.19 (−1.89 to −0.50)	−0.79 (−1.44 to −0.14)	0.017
Catenibacterium mitsuokai et rel.						0.50 (−0.09 to 1.09)	0.31 (−0.26 to 0.87)	0.282
Clostridium leptum et rel.						−0.03 (−0.31 to 0.24)	−0.05 (−0.30 to 0.21)	0.709
Dialister						−0.04 (−0.22 to 0.14)	0.06 (−0.10 to 0.22)	0.441
Eubacterium halii et rel.						0.37 (−0.07 to 0.81)	0.23 (−0.18 to 0.64)	0.267
Eubacterium rectale et rel.						0.04 (−0.36 to 0.45)	−0.04 (−0.41 to 0.33)	0.836
Eubacterium ventriosum et rel.						0.08 (−0.19 to 0.34)	−0.01 (−0.26 to 0.24)	0.915
Parabacteroides distasonis et rel.						−0.32 (−0.68 to 0.04)	−0.17 (−0.50 to 0.16)	0.324
Ruminococcus callidus et rel.						0.12 (−0.21 to 0.45)	0.04 (−0.26 to 0.33)	0.816
UCI						0.19 (−0.08 to 0.45)	−0.00 (−0.24 to 0.24)	0.997
Bacterial groups exhibiting bimodality, n (%)
Prevotella melaninogenica et rel.: Low (<4.46); High (≥ 4.46)	156 (52.9); 139 (47.1)		0 (reference); −0.03 (−0.24 to 0.17)	0 (reference); −0.14 (−0.34 to 0.05)	0.154	0 (reference); 0.05 (−0.17 to 0.27)	0 (reference); −0.07 (−0.27 to 0.13)	0.490
Prevotella oralis et rel.: Low (<3.71); High (≥ 3.71)	154 (52.2); 141 (47.8)		0 (reference); −0.04 (−0.24 to 0.17)	0 (reference); −0.15 (−0.34 to 0.05)	0.148	0 (reference); 0.05 (−0.17 to 0.24)	0 (reference); −0.06 (−0.26 to 0.14)	0.537
Dialister: Low (<3.03); High (≥ 3.03)	165 (55.9); 130 (44.1)		0 (reference); −0.12 (−0.33 to 0.08)	0 (reference); −0.02 (−0.18 to 0.22)	0.862	0 (reference); −0.13 (−0.35 to 0.08)	0 (reference); 0.05 (−0.15 to 0.25)	0.602
UCII: Low (<3.47); High (≥ 3.47)	126 (42.7); 169 (57.3)		0 (reference); −0.19 (−0.39 to 0.02)	0 (reference); −0.22 (−0.42 to −0.02)	0.035	0 (reference); −0.18 (−0.41 to 0.04)	0 (reference); −0.22 (−0.42 to −0.02)	0.031
Modeling associations with richness, diversity, B:F ratio
B:F ratio			−0.01 (−0.07 to 0.05)	−0.03 (−0.10 to 0.31)	0.320	0.02 (−0.04 to 0.09)	0.00 (−0.06 to 0.06)	0.978
Diversity			−0.27 (−0.85 to 0.32)	−0.15 (−0.73 to 0.43)	0.612	−0.38 (−1.01 to 0.25)	−0.26 (−0.85 to 0.32)	0.376
Richnesse			−0.07 (−0.17 to 0.03)	−0.07 (−0.18 to 0.03)	0.145	−0.07 (−0.17 to 0.03)	−0.08 (0.18 to 0.02)	0.131

^aModel 1: Univariate linear regression models with each of the following: selected bacterial group, bimodal group, B:F ratio, diversity, and richness in its own model.

^b β: Regression coefficients (with 95% CI) from unadjusted and adjusted linear regression analysis, respectively.

^cModel 2: Adjusted for recruitment group (conventional or alternative), maternal educational level (low, middle, high, and others), maternal prepregnancy weight, weight gained during pregnancy, maternal smoking status (yes/no), place and mode of delivery (vaginal delivery at home, vaginal delivery in hospital, or cesarean section in hospital), gestational age, sex, birth weight, household size, duration of breastfeeding, dietary intake (total fiber, total energy, fats, and carbohydrates both as percentage of total energy), antibiotic use (no antibiotic use for the past year, antibiotic use > 4 weeks ago, or antibiotic use ≤ 4 weeks ago), physical activity, and age at fecal sample collection.

^dColumn represents P values for adjusted analysis.

^eAnalyzed in its standardized form by dividing “microbial richness” for each child by SD of variable.

Table 3.

Multiple logistic regression results for associations of selected bacterial groups from elastic net, bimodal groups, diversity, richness, and B:F ratio with overweight

		Overweight
		Model 1 a, unadjusted OR (95% CI)	P value d	Model 2 c , adjusted OR (95% CI) b	P value d
Log10 abundance of bacterial groups
Akkermansia		0.24 (0.08 to 0.74)	0.013	0.31 (0.07 to 1.36)	0.119
Sutterella wadsworthensis et rel.		0.16 (0.03 to 0.89)	0.036	0.10 (0.01 to 0.88)	0.038
Bacteroides fragilis et rel.		0.16 (0.05 to 0.56)	0.004	0.46 (0.09 to 2.36)	0.351
Marvinbryantia formatexigens et rel.		0.08 (0.01 to 0.51)	0.007	0.07 (0.01 to 0.73)	0.027
Streptococcus bovis et rel.		2.09 (0.96 to 4.59)	0.065	3.96 (1.34 to 11.74)	0.013
Aeromonas		<0.01 (<0.01 to 0.47)	0.035	<0.01 (<0.01 to 30.1)	0.177
Lactobacillus gasseri et rel.		2.61 (0.75 to 9.09)	0.132	3.43 (0.74 to 16.02)	0.117
Lactobacillus plantarum et rel.		4.19 (0.59 to 30.02)	0.154	7.73 (0.63 to 94.26)	0.109
Bacterial groups exhibiting bimodality: n (%)
Prevotella melanogenica et rel.: Low (<4.46); High (≥ 4.46)	156 (52.9); 139 (47.1)	1 (reference); 0.74 (0.33 to 1.66)	0.465	1 (reference); 0.21 (0.07 to 0.68)	0.009
Prevotella oralis et rel.: Low (<3.71); High (≥ 3.71)	154 (52.2); 141 (47.8)	1 (reference); 0.69 (0.31 to 1.54)	0.361	1 (reference); 0.20 (0.06 to 0.64)	0.006
Dialister: Low (<3.03); High (≥ 3.03)	165 (55.9); 130 (44.1)	1 (reference); 0.54 (0.23 to 1.24)	0.147	1 (reference); 0.72 (0.25 to 2.05)	0.541
UCII: Low (<3.47); High (≥ 3.47)	126 (42.7); 169 (57.3)	1 (reference); 0.48 (0.21 to 1.07)	0.073	1 (reference); 0.28 (0.10 to 0.79)	0.017
Modeling associations with richness, diversity, B:F ratio
B:F ratio		1.00 (0.77 to 1.28)	0.979	0.89 (0.67 to 1.19)	0.431
Diversity		0.16 (0.02 to 1.12)	0.065	0.09 (0.01 to 1.73)	0.111
Richnesse		0.73 (0.49 to 1.06)	0.102	0.64 (0.37 to 1.09)	0.100

^aModel 1: Univariate logistic regression model with each of the following: selected bacterial group, bimodal group, B:F ratio, diversity, and richness in its own model.

^b ${\mathrm{O}\mathrm{d}\mathrm{d}\mathrm{s}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}(\mathrm{O}\mathrm{R})=e}^{\mathrm{u}\mathrm{n}\mathrm{a}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d}(\beta )}\mathrm{o}\mathrm{r}{e}^{\mathrm{a}\mathrm{d}\mathrm{j}\mathrm{u}\mathrm{s}\mathrm{t}\mathrm{e}\mathrm{d}(\beta )}$ (with 95% CI) from the unadjusted and adjusted logistic regression analysis, respectively.

^cModel 2: Adjusted for recruitment group (conventional or alternative), maternal educational level (low, middle, high, and others), maternal prepregnancy weight, weight gained during pregnancy, maternal smoking status (yes/no), place and mode of delivery (vaginal delivery at home, vaginal delivery in hospital, or cesarean section in hospital), gestational age, sex, birth weight, household size, duration of breastfeeding, dietary intake (total fiber, total energy, fats, and carbohydrates both as percentage of total energy), antibiotic use (no antibiotic use for the past year, antibiotic use > 4 weeks ago, or antibiotic use ≤ 4 weeks ago), physical activity, and age at fecal sample collection.

^dColumn represents P values for the unadjusted and adjusted analysis, respectively.

^eAnalyzed in its standardized form by dividing “microbial richness” for each child by SD of variable.

We subsequently examined whether specific bacterial groups were associated with the anthropometric outcomes under study. Elastic net was applied to the 130 genus‐like bacterial groups for each outcome measure separately. A total of 5 (for BMI z scores), 19 (weight z scores), and 9 (overweight) weight‐associated bacterial groups were identified (Supporting Information Figure S2). The elastic net coefficient paths for these bacterial groups with respect to each outcome are presented in Supporting Information Figure S4. The selected bacterial groups were included as continuous variables in the regression models, except for P. melaninogenica et rel., Dialister species (weight z scores), and UCII (overweight) because they were already identified as having a bimodal abundance distribution and, as such, were analyzed as dichotomous variables.

After adjusting for confounders (Supporting Information Table S2; Model 2), BMI z score was inversely associated with Akkermansia (P = 0.030) and Sutterella wadsworthensis et rel. (P = 0.008). Similar results were obtained from Model 3 (Supporting Information Table S2), which included all five preselected genus‐like bacterial groups, other microbial groups (archaea, bimodal bacterial groups), and confounders.

Weight z scores were inversely associated with Akkermansia (P = 0.009), Burkholderia (P = 0.017), and Marvinbryantia formatexigens et rel. (P = 0.050) after adjusting for confounders. In the model with all 19 preselected bacteria adjusted for confounders plus other microbial groups (Supporting Information Table S2; Model 3), only Akkermansia and M formatexigens remained inversely associated with weight z scores.

With regard to overweight, S wadsworthensis (P = 0.038) and M formatexigens et rel. (P = 0.027) were inversely associated with overweight in the adjusted analyses. In contrast, Streptococcus bovis et rel. (P = 0.013) was positively associated with overweight (Table 3; Model 2). In the model with all nine preselected bacteria, other microbial groups, and confounders, only Akkermansia remained significantly inversely associated with overweight (Supporting Information Table S2; Model 3). In a final sensitivity analysis excluding the seven children who had received antibiotics within 1 month prior to fecal sampling, the strength of the associations was retained.

Overall, Akkermansia, S wadsworthensis et rel., and M formatexigens et rel. were consistently associated with the three anthropometric outcomes, even after adjusting for the presence of other gut microbial groups and confounders.

Discussion

Within a population of 295 school‐aged, generally healthy children participating in the KOALA study, we examined the association between microbiota composition and overweight. So far, this is the largest observational, population‐based study on childhood gut microbiota in relation to body weight. Moreover, the extensive data on lifestyle and diet allowed us to carefully control for confounding factors. In this relatively large cohort, the anthropometric outcomes and selected confounders explained very little variation in the overall microbiota composition. Nevertheless, the abundances of several specific bacterial groups were consistently and significantly linked to BMI, weight z scores, and overweight. This supports the idea that, although the effect size of BMI as a microbiota covariate in large populations that are not selected for extreme high or low body weights is generally low, gut microbial composition is associated with anthropometric outcomes.

The gut microbiota of 6‐ to 7‐year‐old children has been less well studied than that of adults, and it is still being debated whether the composition reaches a relatively stable, adult‐like state soon after weaning 32 or continues developing into the teen years 33. Therefore, although an in‐depth comparison with adults is beyond the scope of this paper, we wanted to determine whether the intestinal ecosystem at an age of 6 to 7 years is governed by the same ecosystem properties that drive microbiota configuration in adults. Indeed, we observed several bimodal distributed bacterial groups as well as their co‐correlating species that were previously reported to be present in healthy adults 31, including UCII, Prevotella species (P oralis and P melaninogenica et rel.) co‐correlating with P ruminicola et rel., and Dialister with uncultured Selenomodaceae. Lack of support for bimodality of UCI in our study is also in concordance with Lahti et al. 31, who previously showed that UCI exhibited very clear shifting state probabilities associated with aging, in that the high abundant state was observed only above 40 years of age. Because the bootstrap support of bimodality for B fragilis was only moderate in the study of Lahti et al., this bimodality might have been missed in our study because of the smaller sample size, age, or health status. These results confirm that, although the microbiota composition might not be the same as in adults, the microbiota in 6‐ to 7‐year‐old children share the same drivers for overall ecosystem configuration.

The abundance of the four bimodal groups explained 26.1% of the variation in species composition compared with < 1% for each of the anthropometric outcomes, which is in line with recent microbiota data from ~4,000 adult individuals 34. In this study by Falony et al., microbiota variation was also defined by the highly fluctuating abundance of several dominant microbiota members, such as Prevotella, whereas the cumulative effect size of BMI as a microbiota covariate, using the same stepwise redundancy analysis, was a small but significant 1% compared with 0.5% (P = 0.07) reported in the present study. Hollister et al. also found that BMI did not account for a significant proportion of the variation in gut microbial composition and function of American children aged 7 to 12 years, even though their conclusions were limited by the fact that only a small number of subjects were included 33. These observations are in contrast with the previously reported large‐scale community shifts in mainly rodent studies 35 or individual humans with extreme obesity 36. It seems that in adults and this group of healthy children within a relatively normal weight range, weight and associated parameters are not major drivers of overall microbial composition or vice versa. The B:F ratio was not consistently associated with weight‐related outcomes in the present study. Although previous studies have reported a lower B:F ratio in individuals with overweight or obesity 8, 9, other studies have found the opposite 10, 11 or, analogous to our study, no association at all 12, 13.

A lower richness and diversity of the gut bacterial communities were reported in subjects with obesity and overweight compared with individuals with normal weight 7. While we also observed such a tendency, the association disappeared upon adjusting for confounders. This implies that differences in microbial diversity between lean subjects and subjects with obesity, as observed in previous studies not comprehensively controlling for confounders, may be overestimated. Altogether, our results confirm the results from a meta‐analysis by Walters et al. that the B:F ratio and diversity appear not to be a general feature distinguishing the normal and overweight human gut microbiota across populations 37.

Although we did not find a significant association with overall microbial community composition, we did observe associations between the anthropometric outcomes and the contrasting alternative stable states of specific bacterial groups that could potentially act as tipping elements of the whole gut microbial ecosystem. We confirmed the observation by Lahti et al. that UCII was inversely associated with body weight 32. Higher abundances of P oralis et rel. and P melaninogenica et rel. were inversely associated with overweight. In addition, using the elastic net statistical method of variable selection 28, we identified several bacterial groups, of which Akkermansia, M formatexigens et rel., and S wadsworthensis et rel. were consistently associated with all three anthropometric outcomes. The inverse association between Akkermansia abundance and body weight is consistent with previous studies both in mice and humans 6, 38. Although the precise mechanism through which Akkermansia influences host metabolism has not yet been fully elucidated, studies in mice have demonstrated that it is involved in the reduction of metabolic endotoxemia, which is characteristic of obesity and associated metabolic disorders, by the restoration of gut barrier function through a membrane protein that interacts with Toll‐like receptor 2 39. M formatexigens et rel. has been shown to ferment glucose to acetate in the presence of high formate concentrations 40, and the production of acetate can result in appetite suppression 41, suggesting a mechanism by which these bacteria might be linked to lower weight. S wadsworthensis et rel. has been suggested to a play a role in autism spectrum disorder in children 42, but its role in overweight development has not been reported.

The strengths of this study compared with previous cross‐sectional studies on the gut microbiota and childhood overweight are its large sample size in combination with a broad and high‐resolution interrogation of the whole gut microbial community, as well as extensive adjustment for important confounding variables. This allowed us to select bacterial groups that were associated with body weight, without constraining the target species a priori. In addition, our study is one of the first to examine the association between microbiota composition and body weight in a population of mostly lean, healthy children. As a consequence, differences in microbiota composition were not very prominent in our study yet were comparable with other large adult data sets and thus provide insight into the role of the gut microbiota in the normal developing child.

A limitation of the present study was the transport time for fecal samples, which ranged from less than 1 day to 3 days at ambient to room temperature. This might have affected the measurable diversity and structure of the bacterial communities. However, several previous studies have shown that the microbial diversity and composition of fecal samples is much more affected by interindividual differences and biases in molecular techniques than differences in short‐term storage conditions, including storage up to 2 weeks at room temperature 43. Furthermore, we were able to assess the associations of the bacterial groups with weight outcomes at only one time point. Only a few longitudinal studies have been published so far in this field 14, 44. Therefore, large longitudinal cohort studies that characterize the gut microbiota at multiple time points and collect detailed data on important confounding variables (e.g., mode of delivery, diet, physical activity) are warranted.

In conclusion, weight‐related parameters (BMI and weight z scores and overweight) are not major drivers of microbial composition in the gut of relatively lean, healthy children. Nonetheless, several specific bacterial taxa appear to be consistently associated with weight‐related outcomes. These include species that have previously been linked to body weight (Akkermansia, UCII), as well as novel species, such as S wadsworthensis et rel. and M formatexigens et rel. More detailed information on their functional role in energy metabolism is needed to establish their importance for weight development. Our results provide new avenues to approach the increasing trend of overweight worldwide.

Funding agencies

This study was funded by TI Food and Nutrition, a public–private partnership in food and nutrition research. Partners are key players in the global food industry, leading research institutes, universities, and medical centers. Additional funding for data collection was received by: Netherlands Organisation for Health Research and Development (grant 2100.0090), Netherlands Asthma Foundation (grants 3.2.03.48, 3.2.07.022), Netherlands Heart Foundation (grant 2008B112), Triodos Foundation, Phoenix Foundation, Raphaël Foundation, Iona Foundation, Foundation for the Advancement of Heilpedagogiek, Royal Friesland Foods (currently FrieslandCampina), Netherlands Sugar Foundation, and the Ministry of Economic Affairs. The sponsors had no influence on the analysis and reporting of this study.

Disclosure

The authors declared no conflict of interest.

Author contributions

CM performed the literature search and data processing, isolated DNA, carried out the statistical analyses, and wrote the manuscript. GH performed the literature search and data processing, carried out microbiota analysis using Human Intestinal Tract Chip and statistical analyses, and wrote the manuscript. JP, PS, MM, CT, PD, and IA contributed to the design of the study and collection of the data. EZ and HS contributed in the Human Intestinal Tract Chip analysis of the microbiota. All authors contributed to the writing of the manuscript, critically reviewed and revised the manuscript, and approved the final manuscript as submitted.

Supporting information